Research On Egg Defect Recognition And Detection Based On Deep Learning

Eggs are one of the most common non-staple foods in daily life.China is the largest producer and consumer of poultry eggs in the world.Various defects will inevitably occur in the production and sales stage of poultry eggs.Generally,unqualified poultry eggs are selected through artificial naked eye identification.However,artificial naked eye detection has physiological and empirical limitations,which is prone to missed detection and false detection.In recent years,some enterprises have tried acoustic and machine vision technology for automatic detection,but there are some problems such as cumbersome detection process,single target and low detection accuracy,which are difficult to achieve the ideal detection effect.This thesis designs a set of egg shell defect detection system based on deep learning to realize the automatic and intelligent detection of eggs.The core of the system is to use convolution neural network to realize egg image recognition and defect target detection.Compared with traditional machine vision which depends on image processing,convolution neural network can extract features by itself through convolution operation,which effectively solves the problems of cumbersome traditional visual detection process,single target Low accuracy.The main work of this thesis is as follows:(1)The types of egg defects and the requirements of system detection are analyzed.Understand the types of egg defects and detection needs through field visit and investigation,and collect egg samples to make image data sets.The collected images are preprocessed appropriately,including image segmentation,image filtering,image enhancement and image annotation,so as to meet the input standard of neural network.(2)Convolutional neural network is used for egg defect recognition.Alex Net,vgg16 and Goog Le Net are built for image recognition.According to the characteristics of egg image,an improved Goog Le Net(Goog Le Net-mini)egg image recognition network is proposed.The experiment shows that the detection accuracy of Goog Le Net-mini network for three kinds of poultry eggs(normal,dirty and crack)is 98.43%,97.45% and 95.88%respectively.Compared with the other three networks,Goog Le Net-mini has better accuracy and model generalization.(3)The defects of poultry eggs are detected by using the YOLOv5 target detection algorithm,and the lightweight improvement of YOLOv5 is made based on the Mobile Netv3 network.The lightweight model Mobile Netv3 is used to replace the backbone network in the original YOLOv5.In addition,combined with the image characteristics,the output layer for target detection of different sizes in YOLOv5 is improved.The parameters and model size of the improved YOLOv5-Mobile Netv3 network are compressed by 35%,the Fps value is 70.2,increased by 4.4,the average detection time of a single image is 14.25 ms,the detection m AP value is 97.7,the detection accuracy of all categories is 95.4%,and the recall rate is 96.7%.(4)The egg defect detection system is designed and developed based on Py Qt.The system integrates image acquisition unit,image processing unit and neural network algorithm model,and develops user management module and information management module.The system can run on Windows platform with good interactivity,simple style and simple operation.After testing,the system has stable operation,reliable performance,the detection accuracy and detection speed can meet the actual detection requirements.